03 · What You Need to Know
Why AI-Generated Interviews Are Not Equivalent to Human Qualitative Data
How can generative AI create fictional research participants?
Large language models can generate responses conditioned on instructions describing a hypothetical person's background, circumstances, attitudes, or experiences.
For example, a researcher might describe a fictional secondary-school teacher with ten years of classroom experience and ask an AI system to answer questions about using digital learning platforms.
The system may produce responses containing apparent classroom experiences, concerns about student participation, and reflections on institutional support.
These responses can resemble authentic interview material because generative models produce language using patterns learned during training and information supplied in the prompt.
However, the fictional teacher has not taught a class, experienced a school policy, or participated in an interview.
The generated account is a model output conditioned on the researcher's instructions. It is not testimony from an actual participant.
What makes an interview transcript genuine research evidence?
An authentic qualitative interview transcript represents an interaction that actually occurred between a researcher and a participant, whether conducted in person, by telephone, or through an online communication platform.
The transcript should correspond to the recorded or otherwise appropriately documented interview, subject to the study's transcription and confidentiality procedures.
Its evidentiary value comes from its relationship to an actual research encounter.
That relationship includes the participant's experiences, the questions asked, the context of the interaction, and the researcher's interpretation of the resulting material.
An AI-generated transcript may reproduce the appearance of these features without reproducing the underlying encounter.
Authentic interview transcript
A documented representation of an interview conducted with an actual participant through the study's research procedures.
AI-generated interview transcript
A simulated conversation produced by an AI system, without the corresponding human interview having occurred.
This distinction remains important even when the simulated transcript appears coherent, emotionally convincing, or consistent with existing research literature.
When does generating a fictional interview become research fabrication?
The United States Office of Research Integrity (ORI) defines fabrication as making up data or results and recording or reporting them.
Under this framework, a researcher who generates fictional interview transcripts and reports them as actual participant interviews misrepresents the origin of the research evidence.
For example, suppose a manuscript reports that 25 teachers participated in semi-structured interviews. In reality, only 15 teachers were interviewed, while the remaining ten transcripts were generated by AI.
If all 25 are reported as genuine human participants, the study contains fabricated interview evidence.
The issue is not resolved by making the artificial responses more realistic or ensuring that their themes resemble those found in the authentic interviews.
The reported participant count, data collection process, and empirical foundation of the analysis would remain inaccurate.
The broader issue of AI-assisted research data fabrication applies regardless of whether the invented observations are textual, numerical, or visual.
Can AI-generated participant quotations be especially misleading?
Yes. Direct quotations carry particular evidentiary significance in qualitative research because they are presented as expressions of participants' experiences, interpretations, or perspectives.
A quotation may support a theme, illustrate an analytical distinction, or demonstrate how a participant understood a particular situation.
When an AI-generated statement is attributed to a fictional participant as though it were genuine testimony, the researcher creates a false connection between the quotation and a human source.
Consider a manuscript that includes the following fictional statement:
Hypothetical AI-Generated Quotation
"I stopped using the learning platform because my students were more concerned about internet access than the lesson itself."
This statement could plausibly appear in a study of technology adoption among teachers. However, if an AI system generated it, the quotation does not establish that any actual teacher expressed this experience.
Its apparent relevance to the research question does not make it empirical evidence.
Researchers should never use generated quotations to strengthen a qualitative theme while representing them as authentic participant statements.
Can AI fabricate an entire qualitative research dataset?
Generative AI can produce collections of fictional interview transcripts, participant profiles, observation notes, and apparent qualitative responses.
Such outputs may be formatted to resemble a complete research corpus, including participant identifiers, interview questions, and thematic variation.
However, a collection of plausible transcripts does not establish that the corresponding research activities occurred.
A fabricated qualitative dataset may create misleading evidence concerning:
- The number and characteristics of participants recruited.
- The experiences and perspectives supposedly reported by those participants.
- The frequency or distribution of themes within the reported sample.
- The existence of particular quotations, incidents, or observations.
- The procedures used to collect and analyze the material.
The resulting analysis may appear internally coherent while lacking an authentic empirical foundation.
This is one reason why qualitative research integrity cannot be assessed solely by examining whether the final themes sound plausible.
Why can AI-generated interviews resemble genuine qualitative data?
Generative AI can produce fluent conversational language, including expressions of uncertainty, emotional reactions, and contextual details.
When given an interview guide, it can generate responses that appear to address each question appropriately.
Nevertheless, the apparent realism may partly reflect the researcher's prompting choices and the model's learned representations.
For example, a prompt describing a teacher as resistant to educational technology may influence the simulated participant to emphasize technological difficulties.
The resulting transcript could appear to confirm an assumption that was already embedded in the instructions.
Such responses should not be interpreted as independent evidence that teachers with those characteristics actually hold the expressed views.
The model is responding to a constructed scenario rather than contributing an independently sampled human perspective.
Can AI-generated participants reproduce demographic or cultural differences?
AI systems may generate different responses when prompted with demographic, occupational, geographic, or cultural characteristics.
However, these differences do not establish that the outputs accurately represent the corresponding human populations.
Generated responses may reflect patterns, stereotypes, omissions, or biases present in the model's training and conditioning.
A fictional participant described as belonging to a particular community cannot substitute for a person who has actually lived within that community.
Researchers should be especially cautious about treating generated narratives as evidence of marginalized groups' experiences or populations that are difficult to recruit.
The difficulty of obtaining authentic participation does not make simulated testimony empirically equivalent.
The separate question of whether AI-generated participants can replace real participants in behavioral research requires attention to representativeness, validity, and the scientific claims being made.
When can simulated interviews be used legitimately?
AI-generated interviews may serve legitimate purposes when the study does not misrepresent them as authentic human testimony.
For example, researchers may use simulated responses to pilot an interview guide, test whether questions are understandable, develop interviewer training materials, or evaluate a preliminary coding framework.
These activities can help identify methodological issues before collecting real participant data.
AI-generated transcripts may also be appropriate as the object of a study investigating how language models represent particular social situations or respond to interview prompts.
In that case, the evidence concerns the behavior of the AI system, not the experiences of human participants.
The methodological use should be clearly stated, and the resulting conclusions should remain within the limits of the data actually generated.
Researchers interested in simulating qualitative interviews with generative AI should distinguish exploratory simulation from empirical qualitative inquiry.
Can simulated interviews help researchers develop a coding framework?
Potentially. Fictional transcripts can provide material for testing whether a proposed coding framework is understandable, whether coding instructions are ambiguous, or whether software correctly applies predefined labels.
For example, a researcher might generate practice responses illustrating possible expressions of satisfaction, frustration, or uncertainty.
These examples can support preliminary coder training or software development.
However, codes and themes identified in artificial transcripts do not establish that those themes occur among real participants.
A coding framework developed with simulated examples should be evaluated against authentic data before researchers make empirical claims about the target population.
The methodological value of using AI-generated responses to test a qualitative coding framework lies in preparing and evaluating the analytical procedure, not in replacing the evidence the procedure is intended to analyze.
Can AI assist with genuine interview transcription without fabricating data?
Yes. AI-based speech recognition can help convert recordings of actual interviews into written transcripts.
This is different from asking a generative model to invent an interview that never occurred.
However, automated transcription can introduce errors, including omitted words, incorrect speaker attribution, and plausible phrases that were not spoken.
Researchers should verify consequential transcript content against the original recording or other authorized source records.
AI-assisted transcription should also comply with participant consent, confidentiality requirements, and applicable data-protection obligations.
Uploading identifiable interview recordings to an external AI service may create separate ethical and privacy concerns even when the transcription task itself is legitimate.
Can AI-generated interview data be mixed with authentic transcripts?
Artificial and authentic material may coexist in a research project for certain methodological purposes, but their provenance must remain explicit.
For example, a researcher might compare coding performance on real interview transcripts and a separately generated synthetic test set.
That design could be legitimate if the datasets are clearly distinguished and the comparison is methodologically justified.
However, adding fictional interviews to the empirical sample and reporting them as additional participants would misrepresent the study.
Researchers should distinguish synthetic data from real research observations throughout data management, analysis, and reporting.
Clear labels alone are not sufficient if the subsequent analysis treats artificial observations as though they were independently collected human testimony.
How can researchers verify that interview evidence is authentic?
Authenticity is best supported through appropriate documentation of the research process rather than assumptions about how genuine language should sound.
Depending on the study, relevant records may include recruitment documentation, consent records, interview schedules, authorized recordings, field notes, transcript versions, and documented analytical decisions.
These records should be managed in ways that protect participant confidentiality.
Not every qualitative study requires audio recording, and the absence of a recording does not automatically indicate fabrication. Some research designs appropriately rely on contemporaneous notes or other approved documentation.
Similarly, a polished or unusually coherent transcript does not prove that AI generated it.
Where concerns arise, investigators should examine the available provenance evidence and follow appropriate research-integrity procedures.
The limitations of detecting AI-fabricated research make accurate documentation particularly important.
Who is responsible when AI-generated interviews appear in a research paper?
Human researchers remain responsible for accurately representing the origin and meaning of the evidence they report.
An AI system cannot assume the ethical and scholarly responsibilities associated with research authorship.
The International Committee of Medical Journal Editors emphasizes that authors remain responsible for the accuracy, integrity, and originality of material produced with AI assistance.
However, responsibility for a formal misconduct finding must be established under the applicable institutional or legal framework.
A researcher who knowingly reports fictional interviews as authentic evidence is in a different position from one who discovers an erroneous transcript insertion and promptly corrects the research record.
The broader principle of human accountability for prohibited AI-assisted research actions remains applicable.
